Apps & data pipelines
OntoSmith builds a knowledge graph in Neo4j under an ontology you approve first: a cheap model proposes facts, and Jev checks each one against the passage it came from before anything is written to the graph.
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
Dongbum Lee splits the work into seats. Claude Haiku 5.5 proposes triples in the reference setup, Jev marks each one verified or sends it to a review band, and an optional Claude Sonnet 5.5 escalation seat, off by default, takes hard calls. His own figures, repeated in a Connected Data London post, include 3.8 cents per document over 500 documents and a graph precision of 0.911 after re-auditing with a judge from another model family.
You can have a developer borrow this design to turn long texts into a searchable web of facts. You first approve a simple list of allowed entity types and relationships. A cheaper AI proposes facts from each passage, and an outside AI judge verifies each fact against the original text before saving it.
To test the tool, your developer needs a running Neo4j database and an access key that connects to the AI models. The author found this graph method improved recall on long contracts, but not on short article abstracts. An AI judge can still make mistakes, so unverified claims get placed into a separate review pile.